Predicting molecule synthesizability using machine learning

Abstract

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for predicting synthesis complexity. In one aspect, a method comprises: receiving data characterizing a molecule; processing a model input that comprises the data characterizing the molecule, using a retrosynthesis featurization machine learning model and in accordance with values of a set of retrosynthesis featurization machine learning model parameters, to generate a feature array comprising a plurality of features of a retrosynthesis tree for the molecule; and processing a model input that comprises the feature array comprising the plurality of features of the retrosynthesis tree for the molecule, using a synthesizability prediction machine learning model and in accordance with values of a set of synthesizability prediction machine learning model parameters, to generate a synthesizability score that characterizes a synthesis complexity of the molecule.

1 . A computer-implemented method comprising:
 receiving data characterizing a molecule;   processing a model input that comprises the data characterizing the molecule, using a retrosynthesis featurization machine learning model and in accordance with values of a set of retrosynthesis featurization machine learning model parameters, to generate a feature array comprising a plurality of features of a retrosynthesis tree for the molecule; and   processing a model input that comprises the feature array comprising the plurality of features of the retrosynthesis tree for the molecule, using a synthesizability prediction machine learning model and in accordance with values of a set of synthesizability prediction machine learning model parameters, to generate a synthesizability score that characterizes a synthesis complexity of the molecule.   
     
     
         2 . The method of  claim 1 , wherein the data characterizing the molecule comprises data defining a chemical structure of the molecule. 
     
     
         3 . The method of  claim 1 , wherein the retrosynthesis featurization machine learning model has been trained by operations comprising:
 obtaining a set of training examples, wherein each training example corresponds to a respective training molecule and comprises: (i) a training model input characterizing the training molecule, and (ii) a target feature array comprising a plurality of features of a retrosynthesis tree for the training molecule; and   training the retrosynthesis featurization machine learning model on the set of training examples.   
     
     
         4 . The method of  claim 3 , wherein a plurality of training examples in the set of training examples have been generated by performing operations comprising:
 obtaining a collection of baseline retrosynthesis trees;   generating a plurality of new retrosynthesis trees by programmatically assembling baseline retrosynthesis trees from the collection of baseline retrosynthesis trees;   generating a respective training example for each new retrosynthesis tree, comprising, for each new retrosynthesis tree:
 generating a training model input for the training example based on a training molecule represented by a root node of the retrosynthesis tree; and 
 generating a target feature array for the training example by processing the new retrosynthesis tree to determine the plurality of features of the retrosynthesis tree for the training molecule. 
   
     
     
         5 . The method of  claim 4 , wherein programmatically assembling baseline retrosynthesis trees from the collection of baseline retrosynthesis trees comprises, for each new retrosynthesis tree:
 searching a space of possible combinations of baseline retrosynthesis trees having leaf nodes that each represent a molecule included in a predefined library of precursor molecules to identify the new retrosynthesis tree, wherein the search is performed to optimize an objective function that measures a penalty associated with the new retrosynthesis tree.   
     
     
         6 . The method of  claim 5 , wherein searching the space of possible combinations of baseline retrosynthesis trees further comprises:
 identifying a first baseline retrosynthesis tree comprising a root node and one or more leaf nodes;   identifying a plurality of second baseline retrosynthesis trees having root nodes that each represent a same molecule as a corresponding leaf node of the first baseline retrosynthesis tree;
 determining, for each of the plurality of second baseline retrosynthesis tree, a penalty that would result from attaching the second baseline retrosynthesis tree to the first baseline retrosynthesis tree at the corresponding leaf node of the first baseline retrosynthesis tree; and 
   selecting a second baseline retrosynthesis tree for attachment to the first baseline retrosynthesis tree at each of the corresponding leaf nodes based on the penalties.   
     
     
         7 . The method of  claim 3 , wherein for a plurality of training examples in the set of training examples, the target feature array of the training example is generated by operations comprising:
 processing data characterizing the training molecule using a retrosynthesis planner to generate a predicted retrosynthesis tree for the training molecule; and   processing the predicted retrosynthesis tree for the training molecule to generate the target feature array of the training example.   
     
     
         8 . The method of  claim 7 , wherein processing the plurality of predicted retrosynthesis trees to generate the target feature array of the training example comprises:
 processing each of the plurality of predicted retrosynthesis trees to generate a corresponding feature array; and   aggregating the feature arrays for the plurality of predicted retrosynthesis trees to generate the target feature array.   
     
     
         9 . The method of  claim 3 , wherein training the retrosynthesis featurization machine learning model on the set of training examples comprises, for each training example:
 training the retrosynthesis featurization machine learning model to reduce a discrepancy between: (i) the target feature array of the training example, and (ii) a predicted feature array generated by processing the training model input of the training example using the retrosynthesis featurization machine learning model.   
     
     
         10 . The method of  claim 1 , wherein the synthesizability prediction machine learning model has been trained by operations comprising:
 obtaining a set of training examples, wherein each training example corresponds to a respective training molecule and comprises: (i) a training model input comprising a feature array including a plurality of features of a retrosynthesis tree for the training molecule, and (ii) a target synthesizability score for the training molecule; and   training the retrosynthesis featurization machine learning model on the set of training examples.   
     
     
         11 . The method of  claim 1 , wherein the synthesizability prediction machine learning model comprises a differentiable machine learning model, and wherein the retrosynthesis featurization machine learning model and the synthesizability prediction machine learning model have been jointly trained by operations comprising:
 obtaining a set of training examples, wherein each training example corresponds to a respective training molecule and comprises: (i) a training model input comprising data characterizing the training molecule, and (ii) a target synthesizability score; and   training the retrosynthesis featurization machine learning model and the synthesizability prediction machine learning model to reduce a discrepancy between: (i) the target synthesizability score of the training example, and (ii) a predicted synthesizability score.   
     
     
         12 . The method of  claim 1 , further comprising using the synthesizability score to perform classifier guidance for a generative model that is configured to generate output molecules, wherein the generative model is configured to generate a ligand molecule for binding to a target protein molecule, and wherein using the synthesizability score to perform classifier guidance for the generative model encourages the generative model to generate output molecules that have at least a threshold level of synthesizability. 
     
     
         13 . The method of  claim 1 , further comprising
 receiving a collection of candidate molecules;   generating a respective synthesizability score for each of the candidate molecules; and   selecting one or more of the candidate molecules for physical synthesis based at least in part on the respective synthesizability scores.   
     
     
         14 . A system comprising one or more computers and one or more storage devices storing instructions that are operable, when executed by the one or more computers, to cause the one or more computers to perform operations comprising:
 receiving data characterizing a molecule;   processing a model input that comprises the data characterizing the molecule, using a retrosynthesis featurization machine learning model and in accordance with values of a set of retrosynthesis featurization machine learning model parameters, to generate a feature array comprising a plurality of features of a retrosynthesis tree for the molecule; and   processing a model input that comprises the feature array comprising the plurality of features of the retrosynthesis tree for the molecule, using a synthesizability prediction machine learning model and in accordance with values of a set of synthesizability prediction machine learning model parameters, to generate a synthesizability score that characterizes a synthesis complexity of the molecule.   
     
     
         15 . The system of  claim 14 , wherein the retrosynthesis featurization machine learning model has been trained by operations comprising:
 obtaining a set of training examples, wherein each training example corresponds to a respective training molecule and comprises: (i) a training model input characterizing the training molecule, and (ii) a target feature array comprising a plurality of features of a retrosynthesis tree for the training molecule; and   training the retrosynthesis featurization machine learning model on the set of training examples.   
     
     
         16 . The system of  claim 15 , wherein a plurality of training examples in the set of training examples have been generated by performing operations comprising:
 obtaining a collection of baseline retrosynthesis trees;   generating a plurality of new retrosynthesis trees by programmatically assembling baseline retrosynthesis trees from the collection of baseline retrosynthesis trees;   generating a respective training example for each new retrosynthesis tree, comprising, for each new retrosynthesis tree:
 generating a training model input for the training example based on a training molecule represented by a root node of the retrosynthesis tree; and 
 generating a target feature array for the training example by processing the new retrosynthesis tree to determine the plurality of features of the retrosynthesis tree for the training molecule. 
   
     
     
         17 . The system of  claim 14 , further comprising using the synthesizability score to perform classifier guidance for a generative model that is configured to generate output molecules, wherein the generative model is configured to generate a ligand molecule for binding to a target protein molecule, and wherein using the synthesizability score to perform classifier guidance for the generative model encourages the generative model to generate output molecules that have at least a threshold level of synthesizability. 
     
     
         18 . One or more non-transitory computer storage media encoded with a computer program, the program comprising instructions that are operable, when executed by one or more computers, to cause the one or more computers to perform operations comprising:
 receiving data characterizing a molecule;   processing a model input that comprises the data characterizing the molecule, using a retrosynthesis featurization machine learning model and in accordance with values of a set of retrosynthesis featurization machine learning model parameters, to generate a feature array comprising a plurality of features of a retrosynthesis tree for the molecule; and   processing a model input that comprises the feature array comprising the plurality of features of the retrosynthesis tree for the molecule, using a synthesizability prediction machine learning model and in accordance with values of a set of synthesizability prediction machine learning model parameters, to generate a synthesizability score that characterizes a synthesis complexity of the molecule.   
     
     
         19 . The non-transitory computer readable media of  claim 18 , wherein the retrosynthesis featurization machine learning model has been trained by operations comprising:
 obtaining a set of training examples, wherein each training example corresponds to a respective training molecule and comprises: (i) a training model input characterizing the training molecule, and (ii) a target feature array comprising a plurality of features of a retrosynthesis tree for the training molecule; and   training the retrosynthesis featurization machine learning model on the set of training examples.   
     
     
         20 . The non-transitory computer readable media of  claim 18 , further comprising using the synthesizability score to perform classifier guidance for a generative model that is configured to generate output molecules, wherein the generative model is configured to generate a ligand molecule for binding to a target protein molecule, and wherein using the synthesizability score to perform classifier guidance for the generative model encourages the generative model to generate output molecules that have at least a threshold level of synthesizability.